Compared with the Nation’s largest defense contractors, the creativity and productivity in development of theory, method, algorithms, and software of the tiny R&D firm, SSPI
(Statistical Statistical | adjective Of or having to do with Statistics, which are summary descriptions computed from finite sets of empirical data; not necessarily related to probability. Signal Processing, Inc.), in the field of signals intelligence is nothing short of phenomenal. The SSPI team was comprised of fewer than ten employees, most of whom were trained by Professor Gardner, SSPI’s president and chief scientist, as PhD students and post-doctoral researchers at the University of California, Davis. The entire budget of SSPI over its 25-year lifetime was less than $25M. The company’s huge store of intellectual property was sold (in essence, gifted) in 2011 to Lockheed Martin Corporation at the time of Gardner’s retirement for a small fraction of this sum and an agreement for placement of SSPI’s senior engineers at the Lockheed Martin Advanced Technology Center in Palo Alto, California.
The following images were excerpted from a 2003 SSPI PowerPoint presentation. They provide a concise overview of some of the technology that was developed by SSPI during the first 15 years of its 25-year history from 1986 to 2011. Following this presentation is an in-depth summary of SSPI’s technology development during its final 10 years. Also available herein on Page 6 is an in-depth survey of SSPI’s early work on application of cyclostationarity theory and method to the specialized field of signal interception. This survey was written 5 years after the 1987 publication of the enabling book [Bk2].
Following is an outline of SSPI’s developed technology in the area of communications signal processing specifically for unintended receivers, as of the end of 2010. The great majority of SSPI’s technology is in the form of technical documents and software (scientific/engineering documentation of innovative theory, methodology, algorithms, software implementations of individual signal processing algorithms and data processing systems of such algorithms, and performance-evaluation data and analysis). These theoretical and methodological results were developed without outside funding. The sources of funding include investments-in-kind of the Owner’s labor, SSPI IRAD funding not billed to any clients as an overhead expense, and some SSPI IRAD funding billed as an overhead expense to contracting customers. Following this outline, a more detailed summary of achievements and status is provided.
These six services represent the Government’s projections for SSPI products to be installed in Government Monitoring Service Delivery Systems.
Background & Challenge
The substantial body of knowledge, experience, and technology that is the standard throughout the monitoring Community for producing emitter-location estimates and calculating approximate confidence regions (or percent containment regions) for emitter location estimates was developed specifically for radar emitters that transmit strong signals relatively frequently from the same location, enabling collectors to make many statistically-independent measurements over time and produce many statistically-independent estimates of emitter location.
In contrast, most uplink multi-user communications emitters are mobile, do not remain at the same location for long, transmit weak signals and only sporadically. In addition, legacy Radar location systems were designed for pulsed Radars and extract TDOA/FDOA measurements before processing for location, whereas communications emitters are typically of the continuous-wave type (digital or analog modulation), not pulsed; so, the original justification for basing all processing for location on TDOA/FDOA measurements typically does not apply to communications-emitter location. Furthermore, communications emitters are typically corrupted by comparable strength cochannel interference, which is relatively unusual for Radar emitters, especially at the densities of interferers encountered with communications emitters.
Despite these fundamental differences between the two classes of emitter-location problems, the successes of the technology developed for radars historically provided strong motivation for taking the approach, to the communications emitter problem of more recent interest, of adapting the older technology to the newer problem.
Yet, it is being increasingly recognized, as the nature of the communications emitter detection/location scenarios of interest evolve along with evolving communications technology, that the standard approximate containment regions can be highly inaccurate for communications-emitter scenarios in contrast with the level of accuracy demonstrated for Radar-emitter scenarios.
Moreover, it has been found more recently that adapting technology from radio frequency emission Imaging (developed for various applications including Radio Astronomy Imaging) can result in important improvements in capability, such as higher detection sensitivity, higher precision, and higher spatial resolution of closely spaced cochannel emitters. This breakthrough has demonstrated that more explicit recognition of the fundamental differences between the two classes of emitter-location problems opens the door to more innovative solutions for communications emitter detection/location, confidence calculation and emitter identification. The areas of technology development that are now widely recognized as needing new methodology include
Response to the Challenge: Synthetic Aperture Zooming
Given this widely recognized need for new methodology, SSPI has “gone back to basics” by achieving the following (as of 2010):
The results of this theoretical work are believed to be revolutionary:
SSPI has developed a highly configurable architecture for implementing a general workhorse version of all the above, collectively referred to as the Bayesian Aperture Synthesis Emitter Location (BASEL) processor. The capability of the BASEL processor is reviewed in summary form on page 11.3.2.
HBC Classifier Development
The concept of the HOCS-based Classifier (HBC) system, in which cyclic cumulants were first proposed as uniquely qualified signal features for classification in cochannel interference, was introduced in 1991 by SSPI on a research project funded jointly by the Army Research Office and the National Science Foundation. (HOCS = Higher-Order Cyclostationarity.) SSPI continued to develop the HBC system for several years under contract with another Government agency. In this period, the HBC was made into a viable end-to-end, stand-alone software tool with band-of-interest detection, filtering, and sample rate conversion, HOCS detection and estimation, feature grouping, and automatic signal classification. Release 1.0 of the HBC was delivered to the Government in 1999.
In 1999, HBC development and evaluation continued under a contract with a different Government agency. In this contract the HBC was tested with real-world data sets and software modifications were made to improve its robustness and usefulness. SSPI delivered release 1.2 of the HBC by the end of 1999
High-level capabilities include:
High-level limitations include:
CuHBC Classifier Development
In 2000-2001 SSPI began development under IR&D of the Custom HBC (CuHBC) for environments with negligible cochannel interference. During this period an initial prototype of the classifier in MATLAB and C was found to offer better performance than the original HBC for cases with negligible cochannel interference but with impairments due to channel distortion found in real data. In 2001 SSPI received a subcontract from Science Applications International Corporation (SAIC) to continue the development of the CuHBC and test it on real data sets. Development continued on the CuHBC system through 2006. During this time, the digital PSK/QAM classification subsystem of CuHBC was integrated with two Government customer’s operational communications monitoring systems, under a contract through Raytheon State College.
High-level capabilities include:
High-level limitations include:
Maximum-Likelihood Classifier Development
During the period 2004 – 2006, SSPI worked under two separate subcontracts with Bit-Systems and Raytheon State College (prime contracts with Government customers) to develop maximum-likelihood classification techniques for signals utilizing M-FSK and QAM modulations Delivered algorithms were integrated into operational system upgrades by RSC as part of a massive upgrade performed by BITS
FSK Classification. A general approach has been developed to perform optimum FSK signal-modulation-classification and signal-parameter-estimation based on the maximum-likelihood principle. Two specific algorithms were developed including (1) ML-FSK Baud-Rate Estimation for M-FSK signals in harsh environments (e.g., low SNR, interference, and highly structured data) and (2) ML-FSK Modulation-Index Estimation for M-FSK signals in moderate to harsh signal environments, with emphasis on estimation of low modulation-index values. The performance of each algorithm has been evaluated by Monte Carlo simulation using predominantly synthetic data.
QAM Classification. Similar to the FSK processing, a general approach has been developed to perform optimum digital-QAM signal-modulation-classification and signal-parameter-estimation based on the maximum-likelihood principle. Parameters of interest include symbol rate, constellation order, and constellation configuration.
Digital PSK/QAM Classifier Development
Presently, SSPI is engaged with a customer to develop a classifier specifically to address the monitoring problem of detection and classification of cochannel signals exhibiting any combination of the following modulation types: PSK2, PSK4, PSK8, QAM16, and OQPSK. The system in development includes band-of-interest detection, filtering, and sample-rate conversion, higher-order cyclic moment detection and estimation, feature grouping, parameter estimation, and automatic signal classification. The classifier development is part of a larger project that includes joint Viterbi demodulation of the set of cochannel waveforms that have been detected and classified. A baseline version of the classifier has just been completed in MATLAB, and performance testing with both collected and synthetic data has begun. Work remains to finalize the classifier design and optimize the performance and configuration.
GSM Joint Demodulation
As part of a multi-year program ending in July 2001, SSPI developed and evaluated two broad classes of signal processing algorithms for detection and copy of cochannel and adjacent-channel interfering GSM signals. The first class of algorithms is designed to exploit cyclostationarity of the GSM waveforms and employs Frequency-Shift (FRESH) filtering technology to achieve signal separation prior to demodulation. The second class of algorithms performs joint signal demodulation using a computationally efficient modification of the Viterbi algorithm. Under this program, SSPI also developed and evaluated a GSM RF environment analyzer designed to enumerate received signals from GSM base stations and mobile users, and to estimate key signal parameters required for subsequent signal geolocation and demodulation. Prototype software for all algorithms was delivered at the conclusion of the program.
The GSM Environment Analyzer (GSM-EA) is a software application for estimating GSM signal and propagation channel parameters. The algorithms are designed to perform robust parameter estimation in multipath and dense cochannel environments.
SSPI has developed a family of techniques that use combined FRESH filtering and fractionally spaced equalizer technology to perform low-cost separation of GSM signals in moderate cochannel signal environments. These techniques utilize a training-assisted, iterative block least squares approach, and are designed to jointly exploit spectral redundancy, known training sequences, and the constant modulus property of the GSM signals. Both single- and multiple-sensor realizations have been developed and evaluated.
SSPI has developed a class of Multiple-Input-Multiple-Output (MIMO) joint maximum likelihood sequence estimation algorithms for performing joint demodulation of cochannel GSM waveforms. SSPI has developed and evaluated various techniques to reduce the computational complexity at the expense of moderately increased Bit Error Rate (BER). These cost-reduction techniques include Constrained Spawning (based on known training and data sequences), the Statistical Thinning Algorithm (STA) for enhanced Viterbi survivor reduction, and methods utilizing Per-Survivor Decision Feedback (PSDF).
CDMA Joint Demodulation
Under a multi-year program ending in November 2006 SSPI developed and evaluated signal processing algorithms for RF environment analysis and multi-user detection (MUD) for CDMA 2G and 2.5G waveforms in highly dense cochannel interference, as seen by overhead collectors. A detailed and complete mathematical model of the received signal environment was developed, based on the protocol specifications. A CDMA Environment Analyzer (CDMA-EA) was developed to detect / enumerate base-station sectors and individual users, and to estimate the channel impulse responses, carrier frequencies, and individual Walsh-channel powers. The family of MUD techniques developed under this program did not require prior knowledge of the PN mask/state and incorporated a prioritization scheme able to process targeted subsets of signals. Two specific MUD techniques were developed and evaluated including: (1) B-SIC (Block Successive Interference Cancellation) with relaxed ML joint demodulation, and (2) the Iterative Sign Algorithm (ISA) for computationally efficient identification and demodulation of the initial data block. Under this program, attainable copy performance was characterized by Monte Carlo simulation for a broad range of signal environments (created using a Government furnished signal simulator, CellSim, as well as SSPI’s own environment generator).
Mixed PSK/QAM Joint Demodulation
SSPI is currently developing a generalization of the Viterbi Algorithm (VA) that is capable of jointly demodulating two cochannel PSK/QAM signals exhibiting (possibly) distinct modulation types and modulation parameters (e.g., carrier frequencies, symbol rates, etc.). This development is being performed as part of a two-year non-Government program that began in May 2009. The most challenging aspect of this joint demodulation problem occurs if the symbol rates of the cochannel signals are different. In this case, the conventional VA is not directly applicable. To solve this problem, SSPI has developed a novel time-variable trellis structure, based on VA concepts, to perform optimal joint demodulation of signals with different symbol rates. The demodulation system incorporates initial synchronization and tracking to acquire and maintain lock on signals with unknown and time-varying amplitude, carrier phase and symbol timing. The demodulator also incorporates optional per-survivor decision feedback and state pruning to minimize the computational cost.
Reaching back farther into the early 1980s, in connection with my work on exploiting cyclostationarity to despread a Direct-Sequence Spread Spectrum signal without using the spreading code (eventually published in the paper [JP14]), Professor Herschel H. Loomis of the Naval Postgraduate School (NPS) in Monterey, California, brought to my attention a class of signal processing problems needing solutions which he suggested may be amenable to techniques based on exploitation of cyclostationarity. This class of problems, referred to as Signal Interception, belongs to the broader field of what is called Signals Intelligence. The focus of applications of my work on cyclostationarity was originally commercial communications systems, a natural outgrowth of my pre-doctoral work at Bell Telephone Laboratories. The signals intelligence problems that SSPI addressed involved primarily radio frequency wireless communications systems, both commercial and military, but was distinct from my earlier work in that the problems addressed arose from the perspective of unintended receivers wanting to extract information from received signals that would be of value for gathering intelligence for the purpose of national defense.
This gave rise to a collaboration that led to the expansion of my modest consulting services to commercial industry into an incorporated research and development firm, SSPI (Statistical Statistical | adjective Of or having to do with Statistics, which are summary descriptions computed from finite sets of empirical data; not necessarily related to probability. Signal Processing Inc.), consisting initially of a group of M.S. and Ph.D. students working on thesis projects at the University of California, Davis, and later expanding to include post-doctoral employees from UCD and elsewhere. In parallel with this development, there was the development of an academic group at NPS, under the direction of Professor Loomis. Whereas SSPI focused on the development of theory and method for tackling signals intelligence challenges, the NPS group focused more on evaluating specific techniques suggested by the theory and associated methodology. This synergistic relationship continued into the early 1990s. Following the first workshop on cyclostationarity in 1992, held in SSPI’s hometown of Yountville, California, and co-sponsored by four independent funding agencies, the National Science Foundation and the Offices of Research of the Army, Navy, and Air Force, SSPI’s customer base grew substantially. Although this detracted from the extent of collaboration between SSPI and NPS, the NPS effort on evaluating cyclostationarity-exploiting techniques also continued to grow, as illustrated in the chronological list of NPS thesis projects included below, which covers the 26-year period from 1983 to 2009.
In addition to UCD and NPS, a key role in applications to signals intelligence was also played by the AFIT (US Air Force Institute of Technology), and other institutions as briefly discussed on Page 6.
UCD Thesis Projects
NPS Thesis Projects